The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied URLs and text prompts to generate synthetic media via remote API routes.
Ingestion points: Input parameters including image_url, audio_url, reference_video_url, and prompt described in SKILL.md.
Capability inventory: Invokes the runcomfy CLI for network operations and remote model processing.
Boundary markers: The skill includes a dedicated "Consent & disclosure" section and a "Security & Privacy" section instructing the agent to refuse non-consensual, defamatory, or harmful requests.
Sanitization: The instructions note that the CLI does not shell-expand prompt content, reducing the risk of shell injection.
[COMMAND_EXECUTION]: The skill utilizes the runcomfy CLI to perform operations. The allowed-tools field in the frontmatter correctly limits the agent's shell access to the specific runcomfy command prefix, adhering to the principle of least privilege.
[EXTERNAL_DOWNLOADS]: The skill references the @runcomfy/cli Node.js package for installation. This is an official vendor-owned resource used for the skill's primary functionality.
[CREDENTIALS_UNSAFE]: The skill mentions that the runcomfy CLI stores tokens in ~/.config/runcomfy/token.json with restricted permissions (0600) and supports RUNCOMFY_TOKEN environment variables. These are standard and safe practices for managing API credentials in CLI tools.

face-swap